Upload 12 files

README.md

@@ -1,7 +1,6 @@
 ---
 license: apache-2.0
 tags:
-- RyzenAI
 - object-detection
 - vision
 - YOLO
@@ -43,7 +42,7 @@ You can use the raw model for object detection. See the [model hub](https://hugg
 
 The dataset MSCOCO2017 contains 118287 images for training and 5000 images for validation.
 
-Download COCO dataset and create directories in your code like this:
+Download COCO dataset and create directories like this:
   ```plain
   └── datasets
        └── coco
@@ -62,7 +61,7 @@ Download COCO dataset and create directories in your code like this:
              └── val2017.txt
   ```
 1. put the val2017 image folder under images directory or use a softlink
-2. the labels folder and val2017.txt above are generate by **general_json2yolo.py**
+2. the labels folder and val2017.txt above are generate by **general_json2yolo.py**, you need put these file in to the datasets/coco folder
 3. modify the coco.yaml like this:
 ```markdown
 path: /path/to/your/datasets/coco  # dataset root dir
@@ -115,9 +114,9 @@ for batch in dataset:
 
 ### Performance
 
-|Metric |Accuracy on IPU|
+|Metric |Quantized onnx|
 | :----:  | :----: |
-|AP\@0.50:0.95|0.486|
+|AP0.50:0.95|48.4|
 
 
 ```bibtex

onnx_eval.py

@@ -78,8 +78,10 @@ class DetectionValidator:
             batch = self.preprocess(batch)
 
             # inference
-            outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: batch["img"].cpu().numpy()})
-            outputs = [torch.tensor(item).to(self.device) for item in outputs]
+            # outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: batch["img"].cpu().numpy()})
+            outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: batch["img"].permute(0, 2, 3, 1).cpu().numpy()})
+            # outputs = [torch.tensor(item).to(self.device) for item in outputs]
+            outputs = [torch.tensor(item).permute(0, 3, 1, 2).to(self.device) for item in outputs]
             preds = post_process(outputs)
 
             # pre-process predictions
@@ -95,6 +97,7 @@ class DetectionValidator:
         return stats
 
     def get_dataloader(self, dataset_path, batch_size):
+        # TODO: manage splits differently
         # calculate stride - check if model is initialized
         return build_dataloader(self.args, batch_size, img_path=dataset_path, stride=32, names=self.data['names'], mode="val")[0]
 
@@ -178,6 +181,7 @@ class DetectionValidator:
                                 ratio_pad=batch["ratio_pad"][si])  # native-space labels
                 labelsn = torch.cat((cls, tbox), 1)  # native-space labels
                 correct_bboxes = self._process_batch(predn, labelsn)
+                # TODO: maybe remove these `self.` arguments as they already are member variable
             self.stats.append((correct_bboxes, pred[:, 4], pred[:, 5], cls.squeeze(-1)))  # (conf, pcls, tcls)
 
             # Save

onnx_inference.py

@@ -78,21 +78,21 @@ def make_parser():
         "--model",
         type=str,
         default="./yolov8m_qat.onnx",
-        help="input your onnx model.",
+        help="Input your onnx model.",
     )
     parser.add_argument(
         "-i",
         "--image_path",
         type=str,
         default='./demo.jpg',
-        help="path to your input image.",
+        help="Path to your input image.",
     )
     parser.add_argument(
         "-o",
         "--output_path",
         type=str,
         default='./demo_infer.jpg',
-        help="path to your output directory.",
+        help="Path to your output directory.",
     )
     parser.add_argument(
         "--ipu", action='store_true', help='flag for ryzen ai'
@@ -133,8 +133,10 @@ if __name__ == '__main__':
         im = preprocess(im)
         if len(im.shape) == 3:
             im = im[None]
-        outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: im.cpu().numpy()})
-        outputs = [torch.tensor(item) for item in outputs]
+        # outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: im.cpu().numpy()})
+        # outputs = [torch.tensor(item) for item in outputs]
+        outputs = onnx_model.run(None, {onnx_model.get_inputs()[0].name: im.permute(0, 2, 3, 1).cpu().numpy()})
+        outputs = [torch.tensor(item).permute(0, 3, 1, 2) for item in outputs]
         preds = post_process(outputs)
         preds = non_max_suppression(
             preds, 0.25, 0.7, agnostic=False, max_det=300, classes=None

utils.py

@@ -851,7 +851,7 @@ def build_dataloader(cfg, batch, img_path, stride=32, rect=False, names=None, ra
         imgsz=cfg.imgsz,
         batch_size=batch,
         augment=mode == "train",  # augmentation
-        hyp=cfg,
+        hyp=cfg,  # TODO: probably add a get_hyps_from_cfg function
         rect=cfg.rect or rect,  # rectangular batches
         cache=cfg.cache or None,
         single_cls=cfg.single_cls or False,
@@ -1170,6 +1170,7 @@ class Bboxes:
         assert bboxes.shape[1] == 4
         self.bboxes = bboxes
         self.format = format
+        # self.normalized = normalized
 
     def convert(self, format):
         assert format in _formats
@@ -1576,7 +1577,7 @@ class YOLODataset(BaseDataset):
                 lb["segments"] = []
         return labels
 
-
+    # TODO: use hyp config to set all these augmentations
     def build_transforms(self, hyp=None):
         transforms = Compose([LetterBox(new_shape=(self.imgsz, self.imgsz), scaleup=False)])
         transforms.append(

yolov8m_qat.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3b770e88b358ad24cc60e7b8bbc00b09bb1e0308f65f45cdcea2a1dfc1301077
-size 103874610
+oid sha256:119038397368b01fee9ad8adcc62061babcf2e2dd417be1946d5bfccb07eb65f
+size 103874987